AI Runs on Water: The Global South Pays the Bill

A UN University report published on June 3, 2026, reveals that the environmental costs of generative AI—including water, land, and electricity—are disproportionately borne by the Global South, while wealthy users in the Global North generate the prompts. The report estimates data centers consumed 415 terawatt-hours in 2024, projected to double by 2030, exceeding Japan's total electricity use.

AI Runs on Water: The Global South Pays the Bill Somewhere in the American desert, a building the size of a small town hums in the dark. It has no windows and almost no people. Inside, tens of thousands of processors run hot, churning through the requests of strangers half a world away: a marketing executive in London asking for a punchier subject line, a student in Toronto summarising a textbook, a hobbyist in Sydney conjuring a cartoon dragon for no particular reason. Pipes carry water through the building to keep the silicon from cooking itself. Transformers the size of lorries pull electricity off the grid in quantities that would once have powered a city. The dragon appears on the hobbyist's screen in about four seconds. The cost of producing it does not appear anywhere at all. That invisibility is the point, and it is also the problem. For most of the people typing into a chatbot, generative artificial intelligence feels like the most weightless technology ever invented. There is no exhaust pipe, no smokestack, no spinning meter on the wall. You ask, it answers, and the bill, if there is one, seems to be a few pennies on a subscription. But the bill is real, and it is enormous, and it is being paid in a currency most users never see: water drawn from stressed aquifers, land scraped flat for server halls, electricity wrenched off ageing grids, and a rising tide of toxic electronic waste. The question the technology industry has been remarkably good at avoiding is a simple one. Who, exactly, is footing it? On 3 June 2026, the United Nations University Institute for Water, Environment and Health published a report that tries to answer that question with the kind of hard numbers the debate has mostly lacked. Its title, “Environmental Cost of AI's Energy Use: Carbon, Water and Land Footprints”, is dry. Its findings are not. The report argues that the environmental costs of the AI boom are not only larger than commonly understood, but are being distributed in a way that is profoundly, structurally unjust. The wealthy generate the prompts. Someone else, very often, pays the bill. The timing was pointed. The report landed in the same week as World Environment Day, an annual fixture in the United Nations calendar, and its authors clearly intended the juxtaposition. While the world's environment ministers issued their usual statements, a team of UN scientists was quietly publishing evidence that one of the fastest-growing pressures on the planet's water, land and atmosphere is a technology that most of those ministers were probably using to draft their speeches. The report is not a polemic. It is an attempt at accounting, an effort to put a defensible number on a cost that the industry has been content to leave uncounted, and then to ask what follows once the number is on the table. The Myth of the Weightless Machine There is a stubborn assumption baked into how we talk about software, and it goes roughly like this: bits are cheap, the cloud is somewhere else, and digital things do not have a physical body. Kaveh Madani, director of UNU-INWEH and one of the report's authors, puts the counterargument bluntly. “Though often described as weightless and virtual,” he says, “the reality of AI is profoundly physical.” That physicality starts with electricity. The International Energy Agency, in its landmark “Energy and AI” analysis published in April 2025, estimated that the world's data centres consumed roughly 415 terawatt-hours of electricity in 2024, about 1.5 per cent of global demand. That figure is already growing at around 12 per cent a year, far faster than overall electricity use. The IEA's central projection is that data centre consumption will roughly double by 2030, reaching around 945 terawatt-hours. That is more than the entire current electricity consumption of Japan. The UNU report adopts the same headline number and spells out what it means: were the world's AI-driven data centres a country, they would rank around eleventh in the world for electricity use, sitting behind France and ahead of Saudi Arabia. The IEA's analysis is careful to note that AI is the single most important driver of this surge, and that the United States accounts for by far the largest share of the projected increase, with China following. In the United States, the agency found, data centres are on track to account for nearly half of all electricity demand growth between now and 2030. This is the part that ought to alarm anyone who follows the energy transition. The grid was already straining to decarbonise. Now it is being asked to absorb a vast new load on top of everything else, and that load does not wait politely for clean power to come online. It plugs into whatever is available, which in most of the world still means gas and coal. None of this is hypothetical. The build-out is happening now, in concrete and copper, across Virginia and Texas, across Inner Mongolia and Ningxia, in Ireland and in the Gulf. And here the UNU report makes its first genuinely clarifying move. The public conversation about AI's energy appetite has fixated on training, the months-long, headline-grabbing process of building a large model from scratch. Training is expensive and dramatic, and it makes for good copy. But it is not where most of the energy goes. It Was Never About Training The report's central technical insight is that the day-to-day running of AI models, the part engineers call inference, accounts for somewhere between 80 and 90 per cent of the technology's total energy demand. Training a model is a one-off cost, however large. Inference is what happens every single time anyone, anywhere, uses the thing. And the using has become astronomical. This matters because it reframes the entire problem. If training were the dominant cost, then the environmental footprint of AI would be lumpy and occasional, a series of expensive sprints punctuating long quiet stretches. You could imagine regulating it the way you might regulate a handful of large industrial projects. But inference is not lumpy. It is continuous, ambient and growing without limit, a constant background draw that scales directly with adoption. The more useful AI becomes, the more it is used, and the more it is used, the heavier its footprint, regardless of how cleverly the original model was trained. The cost is not in the building of the machine. It is in the running of it, forever. Consider a single product. The report notes that ChatGPT alone fields on the order of 2.5 billion prompts a day, and that running it consumes something in the region of 383 gigawatt-hours of electricity a year. That is one application from one company. Multiply the logic across the entire ecosystem of chatbots, image generators, coding assistants, search summaries and the AI features now wedged into every productivity suite on Earth, and the scale of the inference problem comes into focus. It is also wildly uneven from task to task. The report draws on research showing that the energy cost of an AI interaction depends enormously on what you ask for. A simple text query is relatively cheap. Generating an image is, by some measures, more than a thousand times more energy-intensive than a basic text-classification task. Producing even a short, high-resolution AI video can require an order of magnitude more energy again, the report putting a single clip at over 415 watt-hours. Even the quiet creep of AI into ordinary web search carries a cost: the report notes that an AI-enhanced generative search can use roughly ten times the energy of a conventional one. The casual user has no way of knowing any of this. The interface is identical. A request that boils a notional kettle and a request that barely warms a teaspoon look exactly the same on screen, and cost the same nothing at the point of use. Mir Matin, another of the report's authors, frames the accumulation problem precisely. “Every prompt, default setting, generated image, video, and query,” he says, “accumulates when multiplied by billions of users.” This is the crux. No single interaction matters. All of them together matter immensely. And because the cost is spread across billions of weightless-seeming moments, it never lands anywhere a user can feel it. The default settings are perhaps the most insidious detail. When a search engine or an operating system switches on an AI feature by default, billions of people begin paying its resource cost without ever choosing to, and without anyone telling them the choice was made. The Thirst Nobody Mentions If electricity is the part of AI's footprint that gets the headlines, water is the part that gets buried. Data centres are thirsty in two distinct ways. First, the servers inside them generate prodigious heat, and many facilities use evaporative cooling, which works by turning water into vapour and letting it drift away into the atmosphere. That water is gone from the local system. Second, and less obviously, the electricity that powers the centres is itself water-intensive to produce, because thermal power plants use vast quantities of water for cooling. Every kilowatt-hour drawn from a coal or gas plant carries an invisible water cost upstream, before a single drop touches the servers themselves. The pioneering work on this hidden cost came from Shaolei Ren, a researcher at the University of California, Riverside, whose 2023 paper bore the memorable title “Making AI Less 'Thirsty'”. Ren and his colleagues calculated that training GPT-3 in Microsoft's state-of-the-art American data centres could have evaporated around 700,000 litres of clean freshwater, and that the figure would have roughly tripled had the training run been done in the company's less water-efficient Asian facilities. To make the number concrete, his team noted that this was comparable to the water used to manufacture hundreds of cars. Crucially, Ren extended the analysis beyond training to the everyday business of answering queries, and projected that global AI demand could be responsible for the withdrawal of between 4.2 and 6.6 billion cubic metres of water in 2027, more than the total annual water withdrawal of a country the size of Denmark several times over. What makes Ren's work so important is not just the figures but the method. Because operators almost never disclose the water consumption of individual sites, he and his colleagues had to infer it from the efficiency of cooling systems, the local climate, and the water intensity of the electricity feeding each facility. The same prompt, run in a cool and hydro-powered region, might cost a fraction of what it costs in a hot, fossil-fuelled one. The footprint, in other words, is not an intrinsic property of the software. It is a property of where and how the software is run, a point that turns out to matter enormously when you ask who ends up paying. The UNU report takes this body of work and pushes the timeline to 2030, arriving at a figure designed to stop the reader cold. By the end of the decade, it estimates, the annual water footprint of AI could reach 9.3 trillion litres. To make that abstraction tangible, the authors compare it to the basic annual domestic water needs of every one of the 1.3 billion people who live in sub-Saharan Africa. The image is deliberate and devastating: a technology marketed in Silicon Valley and consumed in the world's richest cities, drinking, in effect, the daily water of an entire subcontinent that has barely been consulted about its construction. The geography sharpens the injustice. Data centres are frequently sited where land is cheap, energy is abundant and tax incentives are generous, and those conditions often coincide with regions that are already water-stressed. Matin, whose expertise is in exactly this kind of spatial analysis, has pointed to the danger of mapping where data centres are being built against where water is scarce, and finding the two maps overlapping. A facility that evaporates millions of litres a year in a temperate, rain-soaked region is a manageable nuisance. The same facility in a drought-prone basin is a direct competitor with farms and households for a resource there is not enough of. Communities in such places have already begun to push back, querying why a hyperscale operator should be granted the water their own crops are rationed. Land, Carbon and the Mountain of Waste Water and electricity do not exhaust the inventory. The UNU report adds two further footprints that rarely make it into the conversation at all. The first is land. Server farms are not small. The report projects that the physical land footprint of AI infrastructure could exceed 14,500 square kilometres by 2030, an area it likens to roughly twice the metropolitan expanse of Jakarta, home to more than 32 million people. That is land taken out of other uses, reshaped, fenced, paved and wired, often on the rural fringes of communities that gain a handful of permanent jobs in exchange for a permanent neighbour that never sleeps. The footprint extends well beyond the perimeter fence, too, taking in the substations, transmission corridors and access roads that a facility of this scale demands, and in many cases the dedicated power generation built specifically to feed it. The second is carbon. For all the talk of powering data centres with renewables, the grids they plug into remain substantially fossil-fuelled, and the sheer scale of new demand is, in many regions, keeping coal and gas plants running that might otherwise have closed. The report projects that AI-related activity could be responsible for around 400 million tonnes of carbon dioxide equivalent emissions annually by 2030. To offset that volume of carbon, the authors calculate, would require growing on the order of 6.7 billion trees over a decade. The “green” technology, in other words, is leaning heavily on a decidedly un-green energy system. There is a bitter irony in the fact that some of the same companies championing AI as a tool to fight climate change are, through that very tool, adding materially to the emissions driving it. Then there is the rubbish. Artificial intelligence runs on specialised hardware, principally graphics processing units, that becomes obsolete with brutal speed as each new generation outperforms the last. When the chips are retired, they become electronic waste, a category laced with lead, mercury and other hazardous materials. The scale of the looming problem was quantified in a 2024 study published in Nature Computational Science, led by Peng Wang of the Chinese Academy of Sciences, which projected that the rapid expansion of generative AI could create between 1.2 and 5 million tonnes of additional e-waste over the period to 2030. Under an aggressive-growth scenario, the study found, the annual e-waste stream attributable to generative AI could reach 2.5 million tonnes by 2030, the very figure the UNU report cites. As with water and land, the burden of dealing with that waste tends not to fall on the cities that generated the demand. The world's discarded electronics have a long and grim habit of ending up in informal recycling yards across the global South, where they are picked apart by hand, often by people with no protective equipment and no choice, releasing toxins into the soil, the water and the bodies of the workers themselves. The Two-Country Cloud All of this might be merely alarming if the costs and the benefits were borne by the same people. They are not, and this is where the UNU report moves from environmental accounting into something closer to political economy. The capacity to build and run frontier AI is astonishingly concentrated. The report finds that more than 90 per cent of the world's AI-specialised computing capacity sits in just two countries. The picture is corroborated by independent analysis: a study of the global data centre landscape drawn on by Oxford University researchers found that only 32 countries host AI data centres at all, and that the United States and China between them operate the overwhelming majority of the specialised facilities. By contrast, more than 150 nations have no significant domestic AI compute infrastructure whatsoever. The IEA's own figures show the United States accounting for the largest single share of global data centre electricity consumption, followed by China, with Europe a distant third and the rest of the world barely registering. The concentration runs deeper than geography. The same handful of corporations that own the compute also own the leading models, the training data, the cloud platforms on which everyone else builds, and increasingly the energy deals that keep the whole edifice powered. Oxford's analysts have argued that this clustering of compute, talent, data and capital means that even mid-sized economies, let alone poor ones, face barriers to independent frontier development that are close to insurmountable. The result is a world in which AI is not a general-purpose technology that diffuses outward to everyone, the way electricity or the internet eventually did, but a service piped out from a couple of national hubs, on terms set by their owners. Think about what that distribution actually means. The intelligence is manufactured in a tiny number of places, owned by a tiny number of companies, and rented out to the rest of the planet as a service. The economic returns, the share prices, the productivity gains, the strategic advantage, accrue overwhelmingly to those two countries and the firms headquartered in them. But the environmental costs, as the report documents, are not so neatly contained. They leak. The carbon enters a shared atmosphere that warms everyone. The water is drawn from local basins that, increasingly, sit in the very regions least able to spare it. The discarded hardware migrates down the global waste stream to the poorest places on Earth. This is the asymmetry at the heart of the report, and it deserves to be stated plainly. When someone in a wealthy country generates an image, summarises a document or asks a chatbot for advice, the water, land and energy costs of that interaction are being distributed across communities and ecosystems that had no say in the choice and will see little of the benefit. The convenience is privatised. The cost is, in large part, socialised, and socialised on to exactly the populations with the least power to refuse it. It is a near-perfect inversion of the polluter-pays principle that environmental law spent half a century trying to establish: here, the polluter mostly does not pay, and the payer mostly does not pollute. Why Efficiency Will Not Save Us There is a comforting story the industry likes to tell about all this, and it goes like this: the chips keep getting more efficient, the models keep getting leaner, and so the problem will shrink itself out of existence. Every generation of hardware does more computation per watt. Every clever algorithmic trick squeezes more capability from less silicon. Surely, the argument runs, efficiency will win. It will not, and the reason has a name. It is called the Jevons paradox, after the nineteenth-century English economist William Stanley Jevons, who noticed something counterintuitive about coal. As steam engines became more efficient and burned less coal per unit of work, Britain did not use less coal. It used vastly more, because cheaper, more efficient steam power made coal worth using for a thousand new purposes. Efficiency did not curb consumption. It unleashed it. The same logic stalks artificial intelligence, and the parallel is not merely rhetorical. A 2025 paper presented at the ACM Conference on Fairness, Accountability and Transparency, titled “From Efficiency Gains to Rebound Effects”, examined precisely how Jevons' paradox applies to AI, arguing that the efficiency improvements the industry trumpets as environmental wins are systematically reinvested to expand markets, stimulate new demand and drive aggregate resource consumption upward rather than down. When a model gets cheaper to run, it does not get used the same amount more cheaply. It gets used more. Features that were too costly to ship get shipped. AI gets stuffed into products that never had it. The summary, the autocomplete, the always-on assistant proliferate precisely because each one became cheap. The episode that made this concrete for the whole industry arrived in early 2025, when the Chinese firm DeepSeek released a model that matched the performance of far costlier systems at a fraction of the computational expense. The market's first instinct was to assume this would mean less demand for chips. The more sophisticated reading, which gained ground quickly, was the opposite: drastically cheaper AI would mean drastically more of it, everywhere, all the time. Cheaper inference is not a brake on consumption. It is an accelerator. Madani's co-authored framing captures the trap exactly: more efficient and more affordable AI does not mean less consumption, it means more. This is why the report insists that judging AI's sustainability by efficiency metrics alone, or by carbon alone, is a category error. A model that uses half the energy per query but is used twenty times as often has not solved anything. It has made the problem worse while looking, on the relevant dashboard, like progress. The footprint that matters is the total one, and the total one is going up. The rebound effect is not a quirk to be engineered away. It is the structural reason that efficiency, on its own, can never be the answer, and that some external limit, whether regulatory, economic or physical, will eventually have to do the work that efficiency cannot. The Trouble With Measuring Anything If the costs are this large and this skewed, an obvious question follows: why has it taken so long for anyone to say so clearly? Part of the answer is that the numbers are genuinely hard to pin down, and the companies that hold the best data have shown little appetite for sharing it. Operators rarely disclose the water consumption of individual facilities, the energy mix powering them, or the per-query resource cost of their models. Researchers like Ren have had to reverse-engineer estimates from patchy public filings, regulatory disclosures and educated assumptions about cooling systems and grid composition. The result is a literature full of ranges rather than precise figures, and those ranges are routinely weaponised by industry defenders who point to the uncertainty as a reason to wait. The argument is circular and convenient: the companies decline to publish the data, then cite the resulting uncertainty as grounds for inaction, all while the build-out accelerates. Aczel, the report's lead author, locates the deeper hazard in the metrics themselves, warning that the choices which look greenest on a narrow accounting can disguise real costs that a fuller reckoning would expose. Judge AI's sustainability by carbon alone, and you will systematically miss the water and the land and the waste. Her broader point is that the environmental footprint of AI is not a fixed fact of nature. It is shaped, she argues, by infrastructure decisions, by the energy sources chosen, and by how models are designed, which means it can be shaped differently. That is, in its way, an optimistic claim. If the footprint were destiny, there would be nothing to do but despair. Because it is the product of choices, it is open to better ones. The opacity is not accidental. A technology whose costs are invisible to its users and unmeasured by its public is a technology that faces very little pressure to change. The first act of accountability, then, is simply measurement. You cannot govern what you refuse to count, and for the moment the people best placed to count have every incentive not to. What Accountability Would Actually Look Like The UNU report does not stop at diagnosis. It proposes a framework built on six principles: transparency, efficiency by design, equity and environmental justice, lifecycle responsibility, global cooperation, and sustainable use. The list can read as the usual policy boilerplate, but underneath it sits a genuinely radical proposition, which is that the relationship between AI's beneficiaries and its bill-payers should be made visible and then made fair. Transparency comes first because nothing else works without it. If operators were required to disclose, in standardised and audited form, the energy, water and carbon footprint of their facilities and ideally of their models, the entire debate would shift from contested estimates to verifiable fact. Users could, in principle, see the resource cost of a request the way a car displays its fuel consumption. Regulators could site facilities with full knowledge of local water stress. Investors could price environmental risk properly. The information asymmetry that currently protects the industry would begin to close. None of this requires a technological breakthrough. It requires a disclosure regime, and the political will to impose one. In the weeks after the report appeared, something close to that political will began, tentatively, to surface. On 23 June 2026, at London Climate Action Week, the United Nations Secretary-General, António Guterres, launched what he called the AI Environmental Transparency Initiative, a charter inspired directly by the UNU-INWEH findings. It asks every major AI company to do two things: to measure and publicly disclose the full carbon, water and land footprint of its systems, and to commit to powering every data centre with renewable energy by 2030. “No more hidden costs,” Guterres said. “If AI is to help build a better future, it must be honest about what it costs us now.” Madani, whose report had supplied much of the initiative's intellectual scaffolding, called it “a gift” and “an opportunity to be proactive instead of reactive”, returning to the principle that had animated the whole project: “We cannot properly manage what we do not measure.” The significance is real, and so are the limits. The initiative is a voluntary call rather than a binding rule, an invitation to companies to come clean rather than a mechanism that compels them to. It is the sound of the political will clearing its throat, not yet the disclosure regime itself, and the distance between an industry being asked to disclose and an industry being made to is precisely the distance this report has spent its length measuring. Efficiency by design and lifecycle responsibility push the engineering upstream. The e-waste research is instructive here: Wang's team found that extending the working life of AI hardware, and refurbishing and redeploying ageing chips for less demanding tasks rather than scrapping them, could cut projected e-waste dramatically, by more than 40 per cent in some scenarios. The point generalises. A great deal of AI's footprint is the product of decisions, where to build, what to cool with, how long to run the hardware, whether to bolt an AI feature on to a product that did not need one, and decisions can be made differently. Lifecycle responsibility means the firm that profits from a chip's first life is also accountable for its last, rather than letting the carcass become someone else's problem in a recycling yard half a world away. But the principles that carry the real moral weight are equity, environmental justice and global cooperation, because they speak directly to the asymmetry the rest of the report documents. If 90 per cent of the compute and almost all of the profit sit in two countries, while the water stress, the e-waste and the climate impact land disproportionately on the 150-plus nations with no AI infrastructure of their own, then any honest framework has to grapple with redistribution. That might mean siting standards that steer facilities away from water-scarce regions and on to genuinely surplus renewable power. It might mean the wealthy beneficiaries of AI financing water security, grid resilience and proper e-waste recycling in the places absorbing the downstream costs. It might mean giving those nations a real voice in the governance of a technology that is reshaping their environment without their consent. At minimum, it means refusing to pretend the costs are not there. What the report stops short of, sensibly, is pretending that any of this will be easy. Each principle cuts against a powerful commercial interest. Transparency threatens a competitive secret. Lifecycle responsibility threatens a margin. Equity threatens a status quo from which the powerful benefit enormously. A framework is not a mechanism, and the gap between the two is where most well-meaning governance goes to die. But the value of naming the principles is that it makes the trade-offs explicit. It turns a set of invisible, deferred costs into a visible political question, and visible political questions can at least be argued over, which is more than can be said for costs nobody admits exist. Paying the Bill Return, for a moment, to the windowless building in the desert, and to the four-second dragon. There is nothing wrong with wanting the dragon. The case against the hidden bill is not a case against artificial intelligence, which is already delivering genuine value in medicine, science, accessibility and a hundred mundane corners of working life. The case is against the invisibility. A technology this physical, this thirsty and this geographically lopsided should not be allowed to present itself as weightless, because the weightlessness is a kind of accounting trick, and the trick has victims. The deepest finding of the UNU report is not any single number, alarming as the numbers are. It is the structure those numbers reveal: a global system in which the pleasure of generating is decoupled, almost completely, from the pain of providing. The user in London or Toronto or Sydney experiences AI as frictionless because the friction has been exported, to an aquifer in a dry country, to a grid burning fossil fuel to meet demand it never planned for, to a recycling yard where someone breaks apart a dead processor with their bare hands. The friction did not disappear. It moved to where it could not be seen and could not be refused. Building accountability into that relationship means, in the end, putting the friction back where it belongs. It means the price of a prompt, somewhere, somehow, reflecting the water it evaporated and the carbon it emitted. It means the firms reaping the trillion-dollar valuations carrying the cost of the cleanup, the refurbishment and the repair, rather than letting it flow downhill to people who never typed a word into a chatbot. It means measuring honestly, siting responsibly, and granting the communities on the receiving end something they have never been offered: a say. The world is not about to stop using artificial intelligence. The 2.5 billion daily prompts will become more, not fewer, and the rebound effect guarantees that every efficiency gain will be spent on more usage rather than less impact. The only real question is whether the bill will keep arriving, silently, at the doorsteps of people who never ordered anything, or whether the world musters the will to redirect it to the address where the dragon was actually conjured. The report's authors have done the arithmetic, and in the weeks since, a first move has been made: a United Nations Secretary-General asking the industry to come clean. But it remains an asking, not a requiring, a voluntary charter rather than a bill redirected, and the choice about who pays is still, for now, ours to make. It is worth remembering that someone is already paying, and they are not the ones holding the phone. References - United Nations University Institute for Water, Environment and Health UNU-INWEH . “The Environmental Cost of Artificial Intelligence: Carbon, Water, and Land Footprints.” 3 June 2026. https://unu.edu/inweh/collection/environmental-cost-of-AIs-Enrgy-Use-Carbon-water-and-land-footprints https://unu.edu/inweh/collection/environmental-cost-of-AIs-Enrgy-Use-Carbon-water-and-land-footprints - United Nations University. “Rising Emissions, Depleting Water and Vanishing Land: UN Scientists: AI Is Threatening Natural Resources for Billions.” 3 June 2026. https://unu.edu/inweh/news/environmental-cost-of-AIs-Enrgy-use-carbon-water-and-land-footprints https://unu.edu/inweh/news/environmental-cost-of-AIs-Enrgy-use-carbon-water-and-land-footprints - EurekAlert / UNU-INWEH. “Beyond AI's surging energy use: UN details escalating water, land, and CO2 emission consequences.” 3 June 2026. https://www.eurekalert.org/news-releases/1128642 https://www.eurekalert.org/news-releases/1128642 - UN News. “AI's environmental costs threaten water, land and climate.” 3 June 2026. https://news.un.org/en/story/2026/06/1167658 https://news.un.org/en/story/2026/06/1167658 - International Energy Agency. “Energy and AI: Executive Summary.” April 2025. https://www.iea.org/reports/energy-and-ai/executive-summary https://www.iea.org/reports/energy-and-ai/executive-summary - International Energy Agency. “Energy demand from AI: Energy and AI.” April 2025. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai - Pengfei Li, Jianyi Yang, Mohammad A. Islam, Shaolei Ren. “Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models.” arXiv, 2023. https://arxiv.org/abs/2304.03271 https://arxiv.org/abs/2304.03271 - UC Riverside News. “AI programs consume large volumes of scarce water.” 28 April 2023. https://news.ucr.edu/articles/2023/04/28/ai-programs-consume-large-volumes-scarce-water https://news.ucr.edu/articles/2023/04/28/ai-programs-consume-large-volumes-scarce-water - Pengfei Li, Jianyi Yang, Mohammad A. Islam, Shaolei Ren. “Making AI Less 'Thirsty'.” Communications of the ACM. https://cacm.acm.org/sustainability-and-computing/making-ai-less-thirsty/ https://cacm.acm.org/sustainability-and-computing/making-ai-less-thirsty/ - Vinuesa, R., et al. “From Efficiency Gains to Rebound Effects: The Problem of Jevons' Paradox in AI's Polarized Environmental Debate.” Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency. https://arxiv.org/abs/2501.16548 https://arxiv.org/abs/2501.16548 - NPR Planet Money. “Why the AI world is suddenly obsessed with Jevons paradox.” 4 February 2025. https://www.npr.org/sections/planet-money/2025/02/04/g-s1-46018/ai-deepseek-economics-jevons-paradox https://www.npr.org/sections/planet-money/2025/02/04/g-s1-46018/ai-deepseek-economics-jevons-paradox - Wang, P., et al. “E-waste challenges of generative artificial intelligence.” Nature Computational Science, 2024. https://www.nature.com/articles/s43588-024-00712-6 https://www.nature.com/articles/s43588-024-00712-6 - Scientific American. “Generative AI Could Generate Millions More Tons of E-Waste by 2030.” https://www.scientificamerican.com/article/generative-ai-could-generate-millions-more-tons-of-e-waste-by-2030/ https://www.scientificamerican.com/article/generative-ai-could-generate-millions-more-tons-of-e-waste-by-2030/ - TechRepublic. “Global Divide: Only 32 Countries Host AI Data Centers.” 2025. https://www.techrepublic.com/article/news-only-32-countries-host-data-centers-2025/ https://www.techrepublic.com/article/news-only-32-countries-host-data-centers-2025/ - Oxford Martin AI Governance Initiative. “A Blueprint for Multinational Advanced AI Development.” 2025. https://aigi.ox.ac.uk/publications/a-blueprint-for-multinational-advanced-ai-development/ https://aigi.ox.ac.uk/publications/a-blueprint-for-multinational-advanced-ai-development/ - S&P Global. “Global data center power demand to double by 2030 on AI surge: IEA.” 10 April 2025. https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/041025-global-data-center-power-demand-to-double-by-2030-on-ai-surge-iea https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/041025-global-data-center-power-demand-to-double-by-2030-on-ai-surge-iea - United Nations University Institute for Water, Environment and Health UNU-INWEH . “UN Secretary-General Launches AI Environmental Transparency Initiative, Calling on AI Companies to Disclose Carbon, Water and Land Footprints.” 23 June 2026. https://unu.edu/inweh/news/un-secretary-general-launches-ai-environmental-transparency-initiative-calling-ai https://unu.edu/inweh/news/un-secretary-general-launches-ai-environmental-transparency-initiative-calling-ai - Fortune. “'It is time to come clean': UN Secretary General calls out AI companies on their climate impact.” 23 June 2026. https://fortune.com/2026/06/23/un-guterres-ai-climate-impact-disclosure-data-centers/ https://fortune.com/2026/06/23/un-guterres-ai-climate-impact-disclosure-data-centers/ Tim Green UK-based Systems Theorist & Independent Technology Writer Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk https://smarterarticles.co.uk , challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship. His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities. ORCID: 0009-0002-0156-9795 https://orcid.org/0009-0002-0156-9795 Email: tim@smarterarticles.co.uk mailto:tim@smarterarticles.co.uk Listen to the free weekly SmarterArticles Podcast https://www.smarterarticles.fm